Upload
nguyenkhanh
View
233
Download
1
Embed Size (px)
Citation preview
Prison Crowding and Violent Misconduct
Jonathan Kurzfeld∗
November 3, 2017
Abstract
Justice reform has recently become a popular bipartisan topic in U.S. politics, with reduc-ing the burgeoning U.S. prison population as one of the primary goals. The objective of thisresearch is to estimate the causal relationship between prison crowding and violent behavior.Understanding this relationship is crucial to evaluating the costs of having severely overcrowdedprisons as well as the benefits of reducing such crowding. This study exploits exogenous varia-tion in California prison populations, resulting from a Supreme Court mandate to reduce prisoncrowding, to estimate the effect on violence. Using both difference-in-differences and instru-mental variables identification strategies, a significant positive relationship is identified that isrobust to a variety of model specifications. The estimates suggest that reducing prison crowdingby 10 percentage points leads to a reduction in the rate of assault and battery of approximately12% − 15%. These estimates represent the first in the literature to directly estimate the effectof prison crowding on violent misconduct and do so using a quasi-experimental design. Further-more, heterogeneity of point estimates across treated subpopulations is discussed as evidence ofcompositional change associated decreasing prison crowding.
JEL Codes: K14, K42, D03, A12
Job Market Paper
– Special thanks to the Presley Center for Crime and Justice Studies and the University of California Consortium on
Social Science and Law for their generous support of this research. Additional thanks to Richard Arnott, Joseph Cum-
mins, Mindy Marks, Michael Bates, Gregory DeAngelo, and department colleagues for their invaluable contributions
to this work. –
∗Department of Economics, 3123 Sproul Hall, University of California, Riverside; email: [email protected].
Contents
1 Introduction 1
2 Prison Violence 4
3 California Public Safety Realignment 6
4 Compositional Change: A Theoretical and Empirical Oversight 13
5 CompStat Data 17
6 Empirical Strategies and Results 236.1 Difference-in-Differences Strategy (DD) . . . . . . . . . . . . . . . . . . . . . . . . . 266.2 Instrumental Variable (IV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316.3 Tests & Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
7 Conclusion 39
A Theory Appendix 46
B Empirical Appendix 50
1 Introduction
“The degree of civilization in a society can be measured by entering its prisons.”
– Fyodor Dostoevsky –
The issue of interpersonal violence within prisons is endemic to the administration of justice.
The fundamental nature of the problem is evidenced by the fact that it can be observed through-
out history and across numerous cultures, ideologies, and nations. A study in the U.K. found that
roughly half of inmates reported having been both bully and victim while incarcerated (South &
Wood, 2006). Meanwhile a simple web search for the world’s most violent prisons will turn up
numerous lists that include prisons on every continent and in nations that are rich and poor, devel-
oping and industrialized, and of every religious association1. Given the function of prisons within a
system of justice, the pervasive nature of violence is no great surprise. Nevertheless, it is generally
held that violence within the prison setting is not an intended element of the sanctions imposed
by the judiciary. This position has been affirmed on several occasions by the U.S. Supreme Court
and legal scholars have further argued that violence in prison should be viewed as an infringement
on the human rights of those subjected to incarceration (White, 2008). The preceding implies that
a civilized society must take measures to minimize the presence of prison violence, even in the
absence of direct institutional benefits from doing so.
Yet there are clear fiscal benefits to effectively managing prison violence. This is especially true
in the United States, known to be the most incarcerated nation in the world. Recent estimates
have the U.S. spending approximately $80 billion per year on incarceration and as much as $8.2
billion2 on prison healthcare alone (Glaze and Haberman, 2013). Although only a fraction of the
latter cost is a direct result of violence, there are a host of additional costs associated with violence,
from disability for injured guards to the cost of extending sentences for inmates convicted of new
crimes.
Not surprisingly prison violence has drawn a great deal of attention from academic researchers
seeking to identify individual and institutional characteristics that correlate with violent behavior.
Prison crowding is among the most common institutional characteristics examined but is not a
1Some examples of such lists are here and here.2Report by the PEW Trusts and MacArthur Foundation found here.
1
consistent correlate of violence (Franklin, Franklin, & Pratt, 2006). This paper is the first in this
literature to both adopt a quasi-experimental design and directly estimate the relationship between
prison crowding and violent behavior. The two questions posed in this research are, is there a
causal relationship between crowding and prison violence? And, if so, why has previous research
struggled to provide consistent evidence of such a relationship?
The quasi-experimental design used in this paper relies on a court mandated reduction in the
overall level of crowding in California prisons. On May 23, 2011, California was placed under court
order by the U.S. Supreme Court to reduce its prison population to 137.5% of design capacity or
less within a two year period3. This resulted in the enactment of new legislation that drastically
reduced the flow of inmates into California correctional facilities, creating a plausibly exogenous
shock to the levels of crowding in California prisons. This legislation is referred to as California
Public Safety Realignment or simply AB 109 (henceforth as the latter).
The exogeneity of the shock to crowding is critical to identifying a causal relationship between
crowding and violence because, above and beyond the typical omitted variable concerns, there
remains a plausible simultaneity problem in this setting. Suppose that administrators believe the
crowdedness of a facility does lead to more violent behavior from the inmates, and therefore new
inmates are directed away from facilities that have had recent issues with violence. Then more
crowdedness may lead to more violence, but more violence also leads to less crowdedness. This
simultaneity will depress naive estimates of the effect of crowding on violence and therefore cause
a bias towards zero, assuming the true causal relationship is indeed positive.
Two estimation strategies are used to exploit the exogenous variation created by AB 109, both
using monthly observations of 30 California prisons over more than four years. The first is a
straightforward difference in differences strategy, with the added dimension that several treatments
are estimated simultaneously to account for the distinct way in which AB 109 impacted different
types of inmate populations. The second strategy is an instrumental variables approach that uses
the time intensity of the policy’s effect on crowding, in conjunction with differences in the mix of
population types prior to the shock, to predict changes in crowding. Those predictions are then used
in the second stage regression to estimate the marginal effect of crowding on the rate of assaults.
The two approaches provide robust, statistically significant estimates showing a strong positive
3The full text of the decision can be found at http://www.cdcr.ca.gov/News/docs/USSC-Plata-opinion09-1233.pdf
2
relationship between crowding and violence, particularly among security level 2 populations.
It is a widely accepted doctrine among correctional practitioners, as well as philosophers on
the topic, that crowding causes increased violence. Despite the research attention that has been
paid to this topic, the empirical literature has thus far failed to present consistent evidence of
such a relationship, providing null estimates as often as finding any positive correlation. A second
contribution of this paper is descriptive evidence supporting a hypothesis that helps explain the
puzzling disconnect between evidence and conventional wisdom. A summary of the hypothesis
is that changes in the level of crowding in a prison are generally accompanied by changes to
the composition of the population. This can confound estimates because compositional changes in
individual propensities for violence are distinct from the direct effect of crowding that the researcher
wishes to estimate. The presence of such a compositional effect is discussed in Section 4 and formally
presented in a theoretical framework in Appendix A.
The results of the empirical estimation suggest that a ten percentage point decrease in the level
of crowding (e.g. from 160% of capacity to 150% of capacity) is associated with an approximate
12% to 15% decrease in the rate of assault. To grasp the magnitude of this estimate consider that
a 12% decrease in assaults for the entire state prison population would mean 85 fewer assaults per
month in the state of California. In a more specific example, Avenal State Prison’s total inmate
population fell from 5, 766 to 4, 946 between September 2011 and September 2012, and the prison
was designed for less than 3, 000 inmates. Thats about a 27 percentage point decrease in crowding,
from 192% of capacity down to 165%. Avenal is a prison with a relatively low base rate of violence,
approximately half of the statewide mean, yet this decrease in crowding would still be attributed
a monthly decrease of 4.5 assaults.
This paper proceeds with the following structure. The next section provides background and
context on the study of violence in contemporary prisons. A detailed history and description of
California Public Safety Realignment is given in Section 3. Section 4 defines compositional effects
and their implication for the literature and this research. Section 5 provides a description of the
data that is used in this research and the challenges inherent to that data. Section 6 presents the
two empirical strategies and the resulting estimates. A brief conclusion is presented in Section 7.
3
2 Prison Violence
Researchers have long sought to understand and predict the behavior of those who are incarcerated.
As it stands, the body of literature studying inmate misconduct largely exists within the ambit of
sociology, criminology, and psychology. This literature categorizes the determinants of misconduct
into two groups, individual and institutional. Individual level covariates include both inherent
characteristics – such as race, gender, and age – and historical characteristics – such as educational
attainment and criminal record. Institution level covariates can include numerous environmental
factors, a variety of security measures, and population density or “crowding”.
A major challenge in evaluating the existing evidence in the field is inconsistency in the mea-
surement of violence itself. Many studies use dependent variables that aggregate observations
of violent misconduct with drug-related and other forms of non-violent misconduct (Goetting &
Howsen, 1986; Ruback & Carr, 1993; Wooldredge, Griffin, & Pratt, 2001). This implies the restric-
tive assumption that the marginal effect of a covariate is stable across different types of misconduct,
to which other research has shown contradictory evidence (Camp, Gaes, Langan, & Saylor, 2003;
Steiner & Wooldredge, 2013). A 2013 paper by Steiner & Wooldredge presents the best available
evidence on this issue, suggesting that there are significant differences in correlations of most covari-
ates with the different types of misconduct. Accordingly, this paper narrows the focus to physical
violence or the direct threat thereof, referred to in practice as assault and battery4.
Understanding the primary determinants of violent behavior is foundational to improving man-
agement and enhancing safety within correctional institutions (DiIulio, 1990; Bottoms, 1999). This
is true at every level of management and decision making in the correctional system, from the daily
choices of correctional officers in the prison yard up to the strategic planning and policy decisions of
wardens and legislators. Even the architectural design of correctional facilities has been associated
with some forms of misconduct (Morris & Worrall, 2014).
We can hypothesize three relevant motivations for the study of prison violence, one a purely
academic motive and the others policy oriented. The first, an academic desire to better understand
violent behavior, would imply an interest in the causal effect of each potential covariate, individual
or institutional, as well as any interactions between them. The second motive is to accurately iden-
4The legal definition of ‘battery’ is to cause bodily harm to another. The definition of ‘assault’ is the crediblethreat of bodily harm to another.
4
tify high-risk individuals, for which it is sufficient to simply identify correlations between individual
characteristics and the likelihood of violent misconduct5. The last motive is to better assess the
costs and benefits of policy and/or management options, which demands identification of the causal
effect of institutional characteristics. To base such policy decisions on simple correlations, without
consideration for causality, poses a risk of unexpected and potentially adverse outcomes. Hence in
studying the relationship between crowding and violence, causality plays an important role in the
value of the research.
This research aims to enhance existing evidence of a causal link between crowding and violent
misconduct. Existing research often lacks any form of quasi-experiment or other source of exogeneity
by which to make a claim for causality. Some studies estimate multivariate regressions, with
cross-sectional or panel data, and show that crowding is associated with higher rates of violence
(Megargee, 1977; Gaes & McGuire, 1985). Yet other research contradicts that conclusion, such
as a 2003 study by Camp and several coauthors. Despite being one of very few studies that use
individual inmate data, they do not find consistent evidence of a correlation between crowding and
violence (Camp et al., 2003).
Of the research that does exploit some sort of quasi-experimental design, the exogeneity does
not apply to the relationship between crowding and violence. Chen and Shapiro use regression
discontinuity design to exploit the mechanism by which inmates are assigned to different secu-
rity levels, granting “as good as random” variation with respect to security classification but not
crowding (Chen & Shapiro, 2007). Another study has inmates with the same security classification
randomly assigned to lower security level facilities and examines the implications for misconduct,
but this random variation is also with respect to security level rather than crowding (Camp & Gaes,
2005). Additionally, Camp & Gaes point out that the correlation between crowding and violence
is not statistically significant in their research design.
In recent theoretical research on the topic, Blevins, Listwan, Cullen, & Jonson (2010) propose
a theory synthesizing the several disparate theories of what determines violent behavior in prisons.
Blevins et al. acknowledge crowding as the most common “noxious stimuli” repeatedly linked to
prison assaults and overall misconduct. The precept that crowding leads to increased violence is
5In identifying “high risk” inmates, there are notable social justice and equity concerns to be raised if individualsare ascribed differential treatment based on inherent characteristics. Discussion of these concerns is omitted sincethat motivation does not apply to this research.
5
also the prevailing view among practitioners in the field. However, these views seem at odds with
the fact that empirical evidence of such a link is wholly unconvincing. This puzzle motivates the
main objective of this research, to provide evidence of a causal effect of prison crowding on violent
behavior, if such a link does in fact exist. It equally motivates the secondary objective of trying to
understand why it is so challenging to identify a consistent relationship in the data.
3 California Public Safety Realignment
On May 23rd, 2011, the United States Supreme Court upheld a lower court ruling (Brown v.
Plata) which had determined that the level of overcrowding in California prisons was so severe that
the Eighth Amendment rights of prisoners were being systematically violated. The Court ordered
California to reduce its prison population to 137.5% of design capacity on or before June 27th,
2013. Given the prison population at the time of this order, the required reduction amounted to
approximately one quarter of the existing California prison population. According to the California
Department of Corrections and Rehabilitation (hereafter CDCR) population reports6 the total
prison population in California in January 2011 was over 156,000. This number fell to 135,000
in January 2012, and further to 124,000 in January 2013. Although CDCR did not quite achieve
the full reduction demanded by the courts7, the total prison population declined more swiftly and
significantly than any large US prison population has in recent history.
Because the law that resulted from the court mandate was strict sentencing reform, it is im-
portant to first understand the process by which individuals enter the California prison system.
There are two possible channels by which a new admit is referred to a California prison – either
through the courts following conviction or through the parole system for violation of the parolee’s
conditions of parole. In either case, the new inmate is first sent to one of the select prisons known
as reception centers, of which there were nine in 2011. Inmates are typically held for several months
at the reception center awaiting a classification hearing. This hearing determines which security
classification the inmate should be assigned to, after which the same committee selects a suitable
6Available at http://www.cdcr.ca.gov/Reports Research/Offender Information Services Branch/Population Reports.html7The target reduction was later achieved after the implementation of Proposition 47, which changed sentences for
a set of minor drug and theft offenses from felony to misdemeanor
6
prison for the inmate to serve the remainder of their sentence. The inmate is then transferred to
the assigned prison, where they are housed with inmates of the same security classification.
Within a prison, prisoners of one security classification do not generally interact with prisoners
of other security classifications. Each California prison has several different facilities within it and
these are designed to operate independent of one another. Thus a prisoner that is classified as
security level 2 will be held in a level 2 facility, which is completely separate from the housing and
recreational areas of other security levels. Reception center populations and special needs popula-
tions8 are also held in separate facilities. This segregation within prisons creates the possibility for
greater levels of crowding in some areas than in others.
AB 109
California achieved the massive reduction in population mentioned above by introducing Assembly
Bill 109 (AB 109) and signing it into law. The new law took effect on October 1st, 2011, making
two major changes to how inmates are handled in California. First, technical parole violators
who were previously taken back into state custody are thereafter sent to county jails, with a few
exceptions for very serious violent and sexual offenders. This is a shift from the aforementioned
practice wherein such parole violators were remitted to a nearby state prison for a term of up to
twelve months. The second major element of the reform defined a set of non-violent, non-sexual,
non-serious felony offenses for which the sentences were to be served in county jails rather than
state prisons.
AB 109 did not commute any existing sentences and no prisoners held in state prison prior to
implementation of the law were transferred to county jails; the law only changed who would take
custody of new admissions. Despite the intervention being a change in flow rather than stock, the
impact on the prison population was rapid and distinct. This is depicted in Figure 1. Essentially,
California prisons had a high rate of “churn” in their populations and the new law caused a sharp
decrease at the front end of that churn without immediately affecting the back end. Hence the true
nature of the shock is that the “vacancies” left by released prisoners are no longer being filled at
the same rate as they were prior to the new law.
Note that the shock created by AB 109 was selective by nature, designed to target non-violent
8“Special needs” in the CDCR is what would commonly be known as protective custody.
7
1200
0013
0000
1400
0015
0000
State
Pris
on P
opula
tion
Dec08 Jun09 Dec09 Jun10 Dec10 Jun11 Dec11 June12 Dec12Date
Limited to the population of the 30 included prisonsTime Trend of Total CA Prison Population
Figure 1: Total population of male prisoners incarcerated in the state of California, observed monthly. Thevertical line denotes the last observation prior to implementation of AB 109.
8
offenders9 for diversion. The population decrease can therefore be expected to concentrate among
two subpopulations. The first, rather obvious given that they handle all new admissions, is the
entirety of the shock is channeled through the relatively few reception center facilities in the state.
Parole violations made up a significant proportion of admissions prior to AB 109, so the dramatic
impact on reception centers was predictable. After passing through reception centers the residual
impact is concentrated among security level 2 populations. This is because the point system
by which inmates are assigned a security classification makes it unlikely to begin at the lowest
classification (level 1) and also unlikely that any inmate will accumulate enough points for level
3 unless they have a long criminal record or are convicted of a very serious felony. Given that
the law targets lower level felonies, it follows that we should expect to see the secondary effects
focused among the security level 2 populations of California prisons. Rather than a liability, this
selective compositional feature of the shock provides a measure of variation that is important to the
empirical strategies that follow. This will be discussed further in the data and empirical sections.
Figures 2 and 3 show the time trends of the different subpopulations for the entire state. It is
apparent that the largest decreases are indeed among the reception and security level 2 populations
depicted in figure 2. More specifically, the effect on the reception center populations is immediate,
very sharp, and appears to stabilize again rather quickly. The shock to the level 2 population lags
by one month, not surprisingly, then there is a sharp decline and it takes longer (into the eighth or
ninth month) to stabilize at what appears to be a new equilibrium. Figure 3 clearly demonstrates
that the trends of each of the other subpopulations are not sensitive to the implementation of the
new law. In total, averaging over the six months prior to implementation and the first six months
of 2012, the reception center population falls by about 45% and the Level II population falls by
a more modest 12%. There are also some moderate decreases in other subpopulations, but none
exceed 6% and these can just as likely be attributed to long term downward trends.
Reception Adjustment
As mentioned previously, it was not unexpected that AB 109 would greatly reduce reception center
populations. All else constant, this would have resulted in prisons with reception facilities that
9Parole violators may or may not have been originally convicted of a violent offense, however Orrick & Morris(2015) show that technical parole violators are significantly less likely to engage in misconduct that inmates admittedto prison for new offenses
9
1000
020
000
3000
040
000
Popu
lation
Dec08 Jun09 Dec09 Jun10 Dec10 Jun11 Dec11 June12 Dec12Date
Level 2 ReceptionThe vertical line represents the last observation prior to implementation of AB 109
for ''treated'' subpopulationsTime Trends of Specific Populations
Figure 2: This figure shows the trends for the two subpopulations among which the shock from AB 109 wasconcentrated, summed across all 30 California prisons included in this study. The vertical line denotes thelast observation prior to implementation of AB 109.
010
000
2000
030
000
4000
0Po
pulat
ion
Dec08 Jun09 Dec09 Jun10 Dec10 Jun11 Dec11 June12 Dec12Date
Level 1 Level 3Level 4 Adm.Seg.Special Needs
for ''untreated'' subpopulationsTime Trends of Specific Populations
Figure 3: This figure shows the trends for the major subpopulations which were less impacted by AB 109,summed across all 30 California prisons included in this study. The vertical line denotes the last observationprior to implementation of AB 109.
10
010
000
2000
030
000
Total
Leve
l 3 P
opula
tion
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13Date
Prisons with Rec. Center Prisons w/out Rec. Center
Showing the impact of the reception adjustmentTime Trend of Level 3 Populations
Figure 4: This figure shows the impact of the reception adjustment that followed implementation of AB109. The time trends are for sum of security level 3 populations parsed by whether the prison has a receptioncenter facility or not. The vertical line denotes the last observation prior to implementation of AB 109.
2400
026
000
2800
030
000
3200
0Po
pulat
ion
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13 May13Date
Level 2 Level 3
showing impact of 2013 reclassificationTime Trends of Specific Populations
Figure 5: This figure shows level 2 and level 3 population trends through implementation of both AB 109and the new classification system. The vertical lines denote the last observation prior to implementation ofAB 109 and January 2013.
11
were suddenly near or below their design capacity while other facilities, both at other prisons and
within the same prison, remained severely overcrowded. The Reception Adjustment10 (hereafter
RA) was a reclassification of certain reception facilities to serve an alternate subpopulation. This
did not mean that any given prison was no longer a reception center. Each California prison has
between three and nine individually defined facilities within its organizational structure and most
reception centers had two or three of these dedicated to reception populations. So when RA occurs
reception population are simply consolidated into fewer facilities.
RA complicates the impact of AB 109 in two ways. First, it makes the decrease in actual
crowding for reception center populations less dramatic than implied by the population reduction.
Although the statewide reception population is approximately halved by AB 109, to a smaller degree
RA also decreases the design capacity dedicated to that population. Therefore RA diminishes the
degree of the AB 109 shock to crowding in reception populations. Simultaneously, it decreases
crowding for some other population by increasing the total design capacity devoted to them in
the state. In nearly every case the repurposed facilities were filled with level 3 inmates. Although
some of the new inmates were likely transferred from within the same prison, the overall transfers
were statewide11. So while AB 109 does not significantly decrease the statewide level 3 population,
evidenced in Figure 3, it does result in decreased levels of crowding for that population.
The evidence of decreased crowding for level 3 populations is given in Figure 4. Looking only
at the number of level 3 inmates in reception centers, the dashed line shows that there were very
few prior to AB 109 and that number begins to climb shortly after implementation of the law. On
the other hand, the total level 3 population in all other facilities, the solid line, decreases in near
perfect synchronicity with the increases at reception centers. The total decrease in non-reception
center level 3 population is of a magnitude similar to the overall impact of AB 109 on the statewide
level 2 population. Similar figures are provided in Appendix B for each of the other subpopulations,
showing that none have a distinct shift like the one seen here for the level 3 population.
10”Reception Adjustment” is the author’s term and does not represent official CDCR language.11Many reception centers did not have any level 3 population prior to this.
12
Reclassification
This time period of the following analysis is truncated prior to January 2013 because of another
policy adjustment that was implemented around that time. Likely in response to the statewide
decrease in level 2 inmates following AB 109, the CDCR commissioned a review of their classification
system. This review and the resulting adjustment in the classification system changed the point
thresholds for certain security classifications. These changes led to a shift of inmates from level 3
classification to level 2 classification. Figure 5 shows the time trends of statewide level 2 and level
3 populations into 2013 and the impact of this policy change is very evident. It is not clear in the
available data what portion of these inmates were transferred between facilities or if some facilities
were repurposed (similar to the RA repurposing) to accommodate the greater number of level 2
inmates.
Although this second policy shock provides opportunity to explore some interesting research
questions about composition and crowding, the variation is quite different from the that induced
by the AB 109 shock. As such the current research is limited to the months preceding the reclas-
sification.
4 Compositional Change: A Theoretical and Empirical Oversight
In any conception of the sequence of events that leads to a violent assault, individually rational
behavior requires that the decisions of each participant be informed by the characteristics and
choices of the other involved parties. As the saying goes, it takes two to “tango”. Given any
degree of heterogeneity among inmates, this implies that changes in the composition of a prison’s
population can influence the rate of violence in that prison. Since any change to the level of
crowding is likely to be accompanied by a change in the composition of the prison population12, it
is a potentially critical oversight to not consider the effect of compositional change in any study of
prison crowding and violence.
12Increasing or decreasing crowding is most often done by increasing or decreasing the population size, which canbe expected to change the composition of the population unless the selection by which changes are made is as goodas random. Such policy or administrative changes tend to be distinctly non-random, as in the case of AB 109 whichspecifically targeted non-violent criminals.
13
This section provides a description of three simultaneous channels, or mechanisms, by which
prison crowdedness may effect the rate of violence in a prison and then summarizes the implications
for this paper and other empirical work in this area. The three mechanisms represent broad
conceptual channels meant to capture all the reasonable means by which crowding may be correlated
with violent behavior. Appendix A develops a model that formalizes the three mechanisms and
presents relevant derivations.
1. Structural Mechanism: Increased crowding leads to increasingly limited personal space
and more individuals sharing a fixed set of available resources – such as basketball courts,
payphones, and restroom facilities – which in turn leads to a higher frequency of poten-
tially contentious encounters between inmates. Therefore increased crowding can cause each
individual to experience a greater number of potentially violent confrontations with other in-
mates in a given time period. This mechanism is likely to be reflected by a positive association
between crowding and violence.
2. Behavioral Mechanism: Crowding exacerbates the individual’s lack of access to personal
freedoms, amenities, and basic necessities, potentially leading to a behavioral increase in an
individual’s willingness to resort to violent behavior. This, in turn, increases the likelihood
that any given contentious interaction between inmates becomes violent13. Sleep is a classic
example of a basic human need whose deprivation has been shown to increase aggressive
behavior (Kamphuis et al., 2012). This mechanism is also likely to be reflected by a positive
association between crowding and violence.
3. Compositional Mechanism: The manner in which a particular change in the level of
crowding changes the composition of the inmate population, especially with respect to in-
dividual propensities for violent behavior. Depending on the specific policy design, this can
result in a population that is either more or less prone to violent behavior, on average. This
mechanism could therefore result in either a positive or negative association between crowding
and violence.13This mechanism may be viewed as an emotional response wherein the individual becomes more irritable or
unstable, or as a rational response where the individual recognizes a lower opportunity cost associated with violentbehavior. The rational response assumes that violent behavior results in punishment with some positive probabilityand the opportunity cost is then the expected utility of avoiding the implied sanctions. That expected utility isreduced if prison life without sanctions grants access to fewer resources or less freedom due to crowding.
14
The first two mechanisms can be thought of as “pure” crowding mechanisms, representing
direct effects of crowding itself. The compositional mechanism is distinct in that it occurs as a
result of correlation between changes in crowding and population composition, where the actual
impact on the rate of violence is derived from the compositional change rather than the change
in crowding. In addition, the nature of compositional change, and thus the resulting effect of the
compositional mechanism on violence, is dependent upon the case specific process of selection by
which the population is increased or decreased.
In the case of AB 109, as well as nearly any other policy intervention meant to decrease crowding,
the new law decreases crowding by reducing the portion of the population least likely to be prone to
violent behavior. After implementation of the law, offenders convicted of non-violent, non-serious,
and non-sexual felonies are no longer sentenced to serve time in state prisons; this effectively
redirected many of the least violent new admissions away from California prisons14. Thus the
resulting distribution prisoners is expected to be more prone to violent behavior, on average. So
while the structural and behavioral mechanisms may decrease the rate of violence after AB 109,
the compositional mechanism is expected to do the opposite, increasing the rate of violence.
The three mechanisms can be conceptually contrasted with the example of a prisoner using the
payphone in the prison yard. Consider a relatively well behaved inmate, Richard, at a maximum
security prison who calls his mother from the prison payphone every Sunday. AB 109 takes effect
and the prison population decreases. As the prison becomes less crowded, it becomes less likely
that another inmate will seek to use the phone while Richard is using it, harassing him to cut his
call short. This is an example of the structural mechanism since such an interaction could have
resulted in violence. At the same time, Richard gets more sleep now that the prison has reverted
from triple to double occupancy in each cell. He is therefore marginally more patient and less prone
to losing his temper. This constitutes an element of the behavioral mechanism. On the other hand,
the diminished population in the prison is due to fewer non-violent offenders who are less likely to
resort to violence than certain other types of offenders. As a result when Richard is harassed it is
more likely to be by an inmate who is aggressive and prone to violence than prior to AB 109. This
is the effect of the compositional mechanism.
14Research shows that violent crime and past criminal records are strong predictors of future violent misconduct(Walters & Crawford, 2013).
15
The implications of these mechanisms for the empirical study presented in this paper can be
understood through Equation 1. The elasticities in this equation are derived from an identity
(Equation 5 in the appendix) stating that the total violence in a prison will equal the average
probability that any pairwise interaction results in violence multiplied by the total number of such
contentious interactions in the prison per unit time. The full model and derivation is available in
Appendix A.
The term on the left side of Equation 1 is the elasticity of aggregate violence, V , with respect
to a policy shift parameter, λ. The shift parameter is defined such that prison crowding, c, exhibits
unit elasticity with respect to it (a policy that decreases crowding by 10% is represented by a 10%
decrease in λ). The first term on the right side of the equation is the elasticity of violence with
respect to crowding. This term encompasses both of the “pure” crowding mechanisms from above
– structural and behavioral – and represents the impact of crowding that researchers typically seek
to estimate. Yet, without any direct measurement of compositional changes, the latter elasticity
is inevitably confounded with the “compositional elasticity” represented by the final term in the
equation. This last elasticity measures how the probability of any particular encounter becoming
violent, π, responds to the policy represented by λ. Eπ:λ < 0 implies that as the policy increases
crowding the population becomes less prone to violence or conversely if the policy is decreasing
crowding (∆λ < 0) then the remaining population, on average, becomes more prone to violence.
As explained previously, the latter is exactly what is expected in the case of AB 109.
Derived Elasticity of Aggregate Prison Violence
EV :λ = EV :c + Eπ:λ (1)
Equation 1 provides a mathematical representation of what is proposed in the earlier description
of the mechanisms. Assuming there is some validity to the structural or behavioral mechanisms,
it is the case that EV :c > 1 and therefore the rate of violence increases with the level of crowding.
However, when empirical estimates actually represent EV :λ then the desired EV :c is conflated with
Eπ:λ. Although the sign of Eπ:λ depends on policy design, there are many cases where the obvious
expectation is Eπ:λ < 0 and this implies a downward bias in estimates meant to capture the EV :c
relationship.
16
Unfortunately, California prison data on individual inmate misconduct is not currently available.
Nor is there data measuring capacity or rates of misconduct for the individual facilities within
each prison. This means that it is not feasible with existing data to separately identify the two
crowding mechanisms of this model nor to precisely isolate the compositional mechanism. Therefore
the empirical strategies used in this paper do not attempt to directly identify the compositional
effect of AB 109. Instead, because the AB 109 population shock is concentrated among three
distinct subpopulations and at least two of those interact in predictably different ways with the
compositional element of the policy, the potential for a compositional effect is used to analyze
between-group differences in the point estimates. Reflectively, the differences in point estimates
can also be taken as supportive evidence of an existing compositional mechanism. In addition, the
potential presence of the compositional mechanism contributes the implication that the estimated
coefficients are in fact lower bounds on what would be considered the true effect of crowding on
violence.
5 CompStat Data
Panel data used for this research has been taken from the California Department of Corrections
and Rehabilitation (CDCR) CompStat Reports. These contain monthly observations for each
adult correctional facility in the state of California. There are 35 currently operational prisons in
California, five of which are excluded from the analysis. The excluded prisons are either medical
facilities, female detention centers, or were not operational at the time of the policy shock. The
time period of analysis is limited to July 2008 through December 2012, due to irregularities in the
earlier data and the January 2013 reclassification that was discussed in section 3. Given that AB
109 was implemented at the beginning of October 2011, the data in use spans three years prior to
and fifteen months post implementation of the law.
The variable of interest is the level of overcrowding in each prison, which is constructed in this
analysis from two variables in the CompStat data. Total population is a simple count of the total
number of inmates held in a prison in a given month. Design beds is reportedly the number of
beds that a given facility was designed to hold. The institutional definition of design beds for a
17
facility is a single bunk per cell and single level bunks in dorm housing15. However, there is minor
month-to-month variation in the CompStat measure of design beds. This variation is not consistent
with the idea that prisons have a fixed capacity – notwithstanding new construction or demolition,
which would not be characterized by such frequent and minor changes. Therefore design beds for
each prison is averaged over the six months preceding implementation of AB 109 and this is taken
as the fixed design capacity of each prison. Table 1 shows that these two measure of capacity are
quite similar, fixed design capacity having only slightly less variation. Prison crowding (crowdit) is
defined in this research as the ratio of total population to fixed design capacity16.
The CompStat data includes a number of measures of misconduct, broadly defined as either
disciplinaries or incidents. Disciplinaries are individual reports of misconduct for each prisoner,
which are included in their personal files. There are several different types of disciplinary, ranging
from simple conduct or cell phone possession to assault and battery or murder. Incidents are
recorded in slightly more detail (such as the type of drug confiscated) but with less frequency,
suggesting that a disciplinary can be issued without having to write up a full incident report or
that each incident may involve an unspecified number of individuals.
The dependent variable used for this research is the monthly sum of disciplinary reports for
assault on an inmate, assault on staff, battery on an inmate, and battery on staff17. The inclusion
of murder and attempted murder in this sum does not noticeably impact the estimates reported
below, which is expected since such attacks are relatively infrequent. Although assault and battery
are measured separately in the more recent CompStat reports, they were not in the earlier years
and are therefore summed into a single variable, assaults, in the statistics below.
In Table 1 assault statistics are given for all assaults and then broken down by whether the
victim was a staff member or an inmate. Assaults on inmates are the most frequent, averaging more
than 18 per month in each prison; however, at just below 5 per month, assaults on staff members
are also quite common. The other categories of violent misconduct occur with less frequency and
15This definition of capacity could be seen as reasonably conservative, which may explain why the target given inthe court mandate was for the CDCR to reduce overcrowding to only 137.5% of design capacity.
16This measure of prison crowding differs slightly from the official measures used by the CDCR, likely due to themanner in which fixed design capacity is constructed. The measure defined here does closely track the official measureand has been determined to be the most appropriate for this research given the timing of the policy shock beingstudied.
17As defined previously, the legal distinction between an assault and a battery comes down to the credible threatof bodily harm versus actually causing bodily harm.
18
Table 1: Simple Summary Statistics
count mean sd min max
Measures of Violence
Rate of Assaults 1620 .537083 .3496848 0 2.777778
Total Assaults 1620 23.45247 14.4966 0 133
Assaults on Inmates 1620 18.72531 12.75872 0 126
Assaults on Staff 1620 4.72716 4.829611 0 58
Murders and Attempts 1620 .5197531 2.433345 0 67
Rioters 900 10.43556 28.4308 0 466
Resisting Staff 900 3.455556 5.287425 0 49
Posessions of a Weapon 899 5.292547 5.950058 0 46
Population & Crowding
Crowding (P/K) 1620 1.872671 .2652445 1.155107 2.44784
Total Population (P) 1620 4566.968 1084.344 2212 7179
Design Beds 1620 2451.812 612.4285 1234 3880
Fixed Design Capacity (K) 1620 2466.111 608.9742 1557 3789.333
Population Shares
Security Level 1 1620 .1094138 .1547788 0 .685277
Security Level 2 1620 .2131347 .2962029 0 .9995067
Security Level 3 1620 .2329096 .2615549 0 .9088176
Security Level 4 1620 .1967216 .2658214 0 .8865633
Reception Center 1620 .1383862 .2605312 0 .9434235
Special Needs 1620 .2042753 .2238708 0 .9279493
Admin. Segregation 1620 .0543071 .0263553 0 .1202622
ADA Inmates 1620 .0637266 .0553348 .0031173 .4041008
CCCMS Inmates 1620 .1969473 .1090689 0 .4756164
Single Inmate Cells 1620 .0473859 .0853995 0 .4496124
Program Enrollment
Prison Industries 1620 148.5117 162.1272 0 616
Academic 1620 413.923 306.4613 0 1687
Non-PIA Work 1620 2251.761 1171.362 0 5591
Subst. Abuse 1620 103.7352 229.5731 0 1818
Subst. Abuse Waitlist 1620 56.23951 103.6996 0 603
Observations 1620
The dependent and key explanatory variables are indicated in italics.
19
most were not reported in the early years of CompStat, which is noticeable in Table 1 by the
fewer numbers of observations for these types of misconduct. Assaults were selected as the most
appropriate category of violence for this study because, in addition to being a reasonable area in
which to expect responsiveness to crowded conditions, their high frequency indicates they are the
most present threat to the safety of inmates and staff. Further, assaults are a distinct and well-
defined form of misconduct, making their measure less susceptible to misreporting by the prison
staff.
The data also include population counts for each of several subpopulations, which are used to
generate population shares of each. There are six major subpopulations: the four levels of security
classification, special needs, and reception center populations. Each of these populations reside in
separate facilities. Administrative segregation units constitute an additional type of facility, but
these are smaller and serve a more temporary purpose. The security classifications are ranked in
ascending level of security threat, one to four. Security level one prisoners are eligible for housing
units that are not within the prison fences and are also permitted to have jobs in the prison that
allow them relatively free movement within the facilities. Security level two and three require much
closer supervision and are not permitted to be outside secure areas, the major difference between
the two levels typically being dormitory housing vs. cells. Security level 4 is reserved for the most
disruptive and violent prisoners, although a sufficiently heinous crime can result in this classification
without any record of institutional misconduct. The special needs and administrative segregation
populations are held apart from the rest, each for their own reasons; one for long-term protective
custody and the other a temporary punitive or safety measure, respectively. There are several
other subpopulations that are not housed exclusively, such as ADA (Americans with Disabilities)
inmates, single-cell inmates, and the CCCMS (Correctional Clinical Case Management Services)
population. The shares of these are included as controls in the empirical strategies. Finally, the
CompStat data includes counts of inmates with life sentences with and without possibility of parole.
Unfortunately, there is a significant amount of missing data for these two variables so they have
been excluded from the analyses.
All of these data are observed at the prison level. However, AB 109 affects crowding, and thus
misconduct, at the facility level. Therefore observed rates of assault are averages across the facilities
within a prison, which makes the population shares of each subpopulation both necessary to the
20
identification strategy and critical as direct controls for the effect of changing shares. The latter is
due to an observable element of compositional change that occurs across the whole prison. Suppose
a prison has a level 2 facility and a level 4 facility, each of equal capacity and population size. The
AB 109 shock reduces the the level 2 population but not the level 4 population. Then, above and
beyond any differences because of reduced crowding, the rate of violence in this prison will have
increased due to the fact that the population of the level 4 facility makes up a greater share of the
total population and level 4 facilities have much higher rates of violence, on average, than level 2
facilities. Thus without controlling for population shares, the implementation of AB 109 would be
correlated with changes in rates of violence that were not due to the effect of diminished crowding.
There are no demographic controls available in these data (such as age and race statistics
for the prison population) but there is fairly detailed information on program participation in
academic, vocational, work, and drug rehabilitation programs. Tables 1 and 2 include variables
for the degree of enrollment in these programs. Prison Industries (PIA) is a work program that
produces marketable products, with higher skill positions that pay relatively high wages. Non-PIA
work positions are the more common prison positions such as maintenance and food service. Each
of these variables is a simple count of enrollment, except for the substance abuse program for which
both enrollment and waitlist are included.
Table 2 breaks the summary statistics into two periods, pre and post implementation of AB 109.
There was a slight downward trend in total population over the full time period, so the difference
in the population means reported in the table slightly exaggerates the actual policy impact on
population and crowding. Table 2 illustrates that there were moderate decreases in all measures
of assault and objectively larger decreases in total population and crowding. One notable point
in this table is the decrease in design beds, minor though it is. This decrease implies that if one
believes design beds to be a more appropriate measure of capacity than the fixed measure used in
this research, then crowding as currently defined actually exaggerates the true impact of AB 109
on California prison crowding. Exaggerating the impact on crowding in this way is of little concern
since it would only result in attenuation bias, which means the estimated coefficients are smaller
and less significant than they would be otherwise.
Table 2 also highlights the dramatic decrease in reception center populations following AB 109.
Excessive focus on this impact to reception populations is discouraged, since the effect on crowding
21
Table 2: Pre and Post Statistics
Pre Post
Measures of Violence
Rate of Assaults 0.55 0.50(0.35) (0.36)
Total Assaults 25.00 19.43(14.62) (13.36)
Assaults on Inmates 19.95 15.53(12.91) (11.78)
Assaults on Staff 5.05 3.90(4.83) (4.74)
Population & Crowding
Crowding (P/K) 1.95 1.67(0.24) (0.20)
Total Population (P) 4755.79 4076.02(1070.81) (959.11)
Design Beds 2480.08 2378.33(613.68) (603.67)
Subpopulations
Level 1 556.89 481.57(846.82) (821.05)
Level 2 1090.50 944.70(1537.78) (1332.32)
Level 3 1059.96 952.60(1180.91) (938.93)
Level 4 792.95 852.42(1076.58) (1109.62)
Reception 778.33 413.69(1403.34) (1017.92)
Special Needs 915.41 1021.08(1033.05) (1039.39)
Program Enrollment
Prison Industries 154.38 133.25(165.36) (152.53)
Academic 436.19 356.03(334.99) (204.55)
Non-PIA Work 2344.31 2011.14(1219.25) (998.36)
Subst. Abuse 128.54 39.24(263.25) (61.89)
Subst. Abuse Waitlist 51.91 67.62(94.22) (124.48)
Observations 1170 450
Means reported. Standard deviations are in parenthesis.
22
was significantly countered by the reception adjustment. An extrapolation based on the observed
decrease in reception populations and increase in level 3 populations at reception centers suggests
the true decrease in crowding at reception facilities was approximately 40 percentage points, only
slightly greater that the decrease at level 2 facilities. This approximation is further supported by
the population means presented by Table 4 in the next section.
6 Empirical Strategies and Results
The simplest and most common approach to estimating the effect of a policy shock such as AB
109 is a difference-in-differences (DD) strategy. This section begins with a DD approach and then
develops a more sophisticated instrumental variables (IV) identification strategy. The IV strategy
builds on the same source of variation as the DD strategy by better capturing the varying time
intensity of treatment across the treated populations. The dependent variable, Yit, in both the DD
and IV strategies is the natural log of the rate of assaults in prison i during month t18. Results in
Appendix B show there is no meaningful change when the dependent variable is altered to include
other violence, such as murder and attempted murder, or to exclude assaults and batteries on staff
members. The log-linear form19 is an intuitive way of incorporating the expectation of a nonlinear
relationship between crowding and violence. Such expected nonlinearities are a natural conclusion
of the frequent conjecture that “violence begets violence”.
The log-linear form, with the crowdit variable constructed as it is, makes the coefficients reported
below semi-elasticities. These are interpreted as the percent change in the rate of assaults associated
with a percentage point change in crowding. For example, the 0.6 point estimate in column 3 of
Table 3 asserts that a ten percentage point increase in crowding is associated with a six percent
increase in the rate of assaults. The percentage point changes in crowding are with respect to the
percent by which the population exceeds the prison’s design capacity.
Table 3 provides estimates from a basic ordinary least squares model with prison fixed effects.
These estimates give an idea of the baseline correlation between crowding and violence in these
data, absent the quasi-experimental design used in later estimates. The model is estimated using
18Rates of violence in this literature are measured per 100 inmates.19The log-linear form the transforms the dependent variable so that marginal changes are approximations of the
percent change in the original variable. Thus a 0.1 increase in ln(Yit) is and approximate 10% increase in Yit.
23
Table 3: OLS Regression with Prison Fixed Effects
Dependent Variable: Log Rate of Assaults
(1) (2) (3) (4)VARIABLES Trend TimeFE Trend TimeFE
Crowding (P/K) 0.624* 0.483 0.601 0.621(0.306) (0.342) (0.443) (0.470)
Security Level 2 -1.527* -1.724** -0.0466 -0.0920(0.784) (0.776) (1.112) (1.137)
Security Level 3 -0.0361 -0.329 0.119 0.103(1.021) (0.989) (1.247) (1.235)
Security Level 4 -0.101 -0.442 0.113 0.0402(1.262) (1.218) (1.296) (1.263)
Reception Center -0.131 -0.377 0.338 0.168(0.797) (0.792) (0.926) (0.927)
Observations 1,440 1,440 1,140 1,140Controls X X X XPeriod Full Full Pre-AB109 Pre-AB109
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
a flexible time trend or time fixed effects and the final two columns restrict the time period to only
include observations prior to implementation of AB 109. The coefficients on crowding are only
marginally significant if at all and suggest a semi-elasticity of approximately 0.5 − 0.6. The weak
statistical significance of these correlations is consistent with the overall state of existing empirical
evidence on this topic, as discussed in Section 2.
The coefficients for the shares of major subpopulations are also reported in Table 3. These
coefficients preview the importance of the share of security level 2 population and its dependence
upon the variation that occurs as a result of AB 109. Note that because the shares of these
subpopulations are generally very stable over time, most of the correlation between assaults and
each share is absorbed by the prison fixed effect. Table 9 in Appendix B shows the raw correlations
between each population share and the rate of assaults. The coefficients in Table 3, on the other
hand, pick up the effects, or lack thereof, of time variation in the shares. However, AB 109 is
responsible for the bulk of such variation and, among the treated populations, the effect of this is
only apparent in the level 2 coefficient. Furthermore, even the effect for level 2 share dissipates
24
-.6-.4
-.20
.2C
hang
e in
the
Rat
e of
Ass
aults
-40 -30 -20 -10 0Change in Percent Beyond Capacity
Control Group Level 2 Group Level 3 Group Reception Group
Excluding reception outlier (DVI)Post-AB 109 Change in Assaults by Change in Crowding
Figure 6: This figure shows the correlations between change in crowding and the change in the rate of assaultfollowing implementation of AB 109, separated by “treatment groups”. Changes are calculated between the6 month average just prior to AB 109 (Apr11 – Sep11) and first half of 2012 (Jan12 – Jun12). The outlierobservation for Deuel Vocational Institute (a member of the reception group) has been omitted from thisfigure.
entirely in columns 3 and 4 when the months following AB 109 are excluded from the regressions.
This suggests that implementation of AB 109 is the source of the only significant variation in the
rate of assaults that is correlated with changing population shares.
Figure 6 emphasizes the importance of the the treated subpopulations in identifying the impacts
of AB 109. The figure depicts the policy induced variation that is exploited by the empirical
strategies in the remainder of this section, the change in prison crowding and the change in the rate
of assault. For the figure, these changes are calculated as the average in the six months preceding
AB 109 differenced from the average in the 4th through 9th months following AB 109. The prison
25
groupings depicted are defined as any prison whose total population is made up of at least 20% of
the given subtype and the control group is the set of prisons that have less than 20% share of each
of the treated subpopulations. Note that these groups are not mutually exclusive. Also, there is
one observation point excluded from Figure 6. One prison in the reception group is such a outlier
that it radically distorts the fitted line for that group. That prison was excluded from the figure for
presentational purposes. However, an otherwise identical figure with the outlier included is shown
in Section 6.3 with a brief discussion of the issue.
What Figure 6 illustrates is the overall impact of AB 109 on crowding and associated changes
in violence in relation to the treated subpopulations. Reductions in crowding are much greater for
all of the treatment groups relative to the control group and those reductions are clearly associated
with reduced violence. In addition, the gradient by which violence changes is notably steeper for
the treatment groups than for the control group.
6.1 Difference-in-Differences Strategy (DD)
Yit = β0Postt + β1Postt ∗ Treati + β2Xit + δi + εit (2)
The concept of the difference-in-differences (DD) strategy in this setting is for a subset of
prisons to be “treated” with an exogenous shock that reduces the degree of crowding in those
prisons, while other prisons remain untreated. In such a case, differencing the post-shock change
in violent behavior for the treated prisons with that of the untreated prisons provides an unbiased
estimate of the causal effect of crowding on violence. The reality of the quasi-experiment provided
by AB 109 deviates from such a straightforward DD setting in two important ways. First, there
are three different subpopulations for which crowding is significantly decreased by AB 109. Each
of these subpopulations experience a different degree or form of treatment and therefore must be
accounted for as three simultaneous but distinct treatments. Second, the unit of observation in the
data, a prison, is not the same as the treatment unit, a facility. This means that each observational
unit is subject to partial treatment by each of the three treatment types, although no one prison
has a large proportion (> 5%) of more than two treated populations.
The basic form of the DD estimation is represented in equation 2. The variable Postt is an
indicator for whether the observation is post-implementation of AB 109. Treati is a vector of
26
variables for the three treatment groups, Reception, Level 2, and Level 3; each is equal to the share
of prison i’s total population that is of the given type, averaged over the six months immediately
preceding implementation of AB 109. Xit is a vector of control variables, which at a minimum
includes indicators for when the reception adjustment occurs at a given reception center. In the
full set of controls (used for most specifications and indicated in the tables by an X in the row
labeled controls) are all variables from Table 1 listed under “Population Shares” and “Program
Enrollment”. δi and εit are the prison fixed effect and an iid error term, respectively. Columns 3
and 4 of Table 5 also include either time fixed effects or a flexible time trend in the specification.
This estimation strategy relies on the identifying assumption that E(εit|t, T reati, Xit) = 0.
The intuitive interpretation of that assumption, as it pertains to AB 109, is that any systematic
variation in the rate of violence pre- and post-implementation, beyond that induced by the shock
to crowding, is uncorrelated with the pre-implementation shares of the treatment populations.
However, the assumption in this case must allow for at least some minimal control variables, Xit,
which is necessary to account for the reception adjustment and the fact that changing population
shares are correlated with implementation of AB 109.
Table 4 provides a view of the data broken down before and after treatment for each of the
different treated populations. The groups are the same as those in Figure 6 and thus do not align
directly with the model in Equation 2, since the latter uses shares rather than a discrete indicator
of treatment. Table 4 illustrates a number of things about the different treatments. All three
treatments see a decrease in the average rate of assault in the post periods, but only marginally so
for the reception group. In addition, there are notable differences in the baseline rates of assault
between the groups, which follow a predictable pattern. Assaults are most common in the control
group because most of the state’s security level 4 population is incarcerated in those prisons.
Reception centers are also prone to higher levels of misconduct, ostensibly because the perpetual
turnover is disruptive to mechanisms of informal governance. Average differences in the rates of
assault across these groups are generally stable over time, with the exception that reception centers
do have greater time variation than other populations.
The highlighted row in Table 4 for crowding shows that all three treatment groups experience
a similar decrease in crowding and it is much larger than the decrease in the control prisons.
This further supports the earlier claim that the reception adjustment was a sufficient response to
27
Table 4: Pre and Post Statistics by Treatment Group
Control Group Reception Group Level 2 Group Level 3 GroupPre Post Pre Post Pre Post Pre Post
ViolenceRate of Assaults 0.71 0.71 0.70 0.66 0.32 0.27 0.51 0.40
(0.41) (0.44) (0.29) (0.32) (0.21) (0.23) (0.32) (0.24)Total Assaults 27.94 25.17 34.79 27.47 17.89 13.01 22.55 15.08
(14.62) (14.30) (14.76) (13.76) (12.02) (11.41) (13.45) (8.53)Assaults on Inmates 21.81 18.97 27.39 23.12 14.61 10.69 17.95 12.01
(13.06) (12.28) (13.80) (13.18) (10.13) (9.82) (12.27) (7.74)Assaults on Staff 6.13 6.20 7.40 4.34 3.28 2.32 4.60 3.07
(4.96) (7.59) (5.25) (2.81) (3.49) (2.51) (5.20) (3.41)
CrowdingCrowding (P/K) 1.77 1.61 2.06 1.70 1.96 1.66 1.99 1.68
(0.25) (0.23) (0.21) (0.20) (0.20) (0.21) (0.20) (0.17)Total Population (P) 4132.62 3740.85 5065.20 4213.28 5493.77 4671.10 4691.94 3985.47
(816.58) (670.22) (684.79) (817.66) (1057.11) (970.41) (1084.07) (1040.11)Design Beds 2390.71 2288.49 2462.17 2395.36 2853.50 2749.36 2404.84 2286.36
(609.32) (568.37) (414.08) (417.62) (623.83) (591.01) (609.38) (643.45)
SubpopulationsLevel 1 718.81 619.01 580.73 521.27 607.61 548.91 352.21 263.08
(1133.89) (1043.51) (624.99) (720.92) (948.91) (909.85) (264.07) (220.14)Level 2 219.69 137.55 496.30 571.23 2907.98 2540.53 842.06 706.93
(357.78) (319.39) (653.22) (745.82) (1405.96) (1147.00) (1148.23) (1032.20)Level 3 298.98 175.03 236.28 622.53 976.08 920.69 2432.76 2040.53
(336.73) (271.48) (327.86) (530.68) (1171.54) (1025.45) (773.24) (557.59)Level 4 2183.19 2158.75 356.33 450.18 230.17 190.08 423.16 533.76
(1165.56) (1125.70) (571.30) (927.09) (578.40) (572.22) (461.50) (514.26)Reception 72.35 26.65 2855.43 1528.01 372.41 118.45 174.23 56.50
(178.73) (87.92) (1212.16) (1481.51) (767.12) (315.42) (551.91) (232.59)Special Needs 683.94 628.47 578.55 906.38 1302.55 1421.83 907.52 1076.85
(653.53) (614.05) (826.60) (1054.63) (1177.03) (1249.39) (1115.88) (1059.35)
ProgramsPrison Industries 22.44 7.80 167.78 122.53 264.08 230.67 242.29 221.54
(38.10) (14.02) (104.32) (77.21) (185.10) (172.39) (190.25) (176.44)Academic 406.55 340.47 172.55 162.53 639.61 480.46 480.33 401.21
(222.53) (141.21) (184.05) (134.68) (391.83) (199.87) (286.15) (180.21)Non-PIA Work 1997.30 1724.30 1346.78 1195.29 3364.15 2839.77 2536.09 2177.30
(1014.45) (921.60) (818.91) (617.45) (1096.89) (875.37) (882.35) (784.79)Subst. Abuse 13.10 0.00 94.13 32.35 294.83 104.55 92.94 30.92
(54.45) (0.00) (137.12) (56.70) (379.89) (58.44) (136.25) (50.87)Subst. Abuse Waitlist 8.50 0.00 67.12 72.66 121.84 156.62 42.87 67.18
(37.34) (0.00) (117.82) (159.85) (113.53) (115.08) (85.39) (118.51)
Observations 273 105 312 120 390 150 429 165Number of IDs 7 7 8 8 10 10 11 11
Means reported. Standard deviations are in parenthesis. Treatment groups defined by > 20% share of the given population.
28
diminish the reception center impact of AB 109 to a level similar to that in level 2 facilities and
simultaneously create a similar magnitude shock to crowding in level 3 facilities. The effect on level
3 populations is also visible in the other highlighted sections of the table, which show the level 3
population experience an approximately equivalent increase and decrease in the reception group
and level 3 group, respectively.
To review, Table 4 and figures in Section 3 demonstrate that each of the three AB 109 treat-
ments provide the crucial variation in crowding necessary to identify the relationship with violence.
Meanwhile the compositional mechanism proposed in Section 4 implies a form of omitted variable
bias for any estimates relying on these reductions in crowding. However, it is further implied that
the expected bias in the level 2 treatment should be minimal, or possibly even absent, while the
expected bias for the reception treatment is larger and potentially quite significant. Minimal bias
for the level 2 treatment relies on similarities between the the base population and that which is
targeted by AB 109, which requires some efficacy to the selection by which security classification
is determined (evidence of this is clear in Table 9). The implication for the level 3 treatment is
not immediately obvious since nothing is known about the selection process by which prisoners
were chosen for transfer to repurposed reception facilities. Indeed it is quite possible that very
different selection criteria were used by officials at different prisons, as opposed to the very uniform
selection criteria that AB 109 implemented for reducing the overall population. Uncertainty about
the exact form of selection in the level 3 treatment indicates that the estimates for this group will
not be particularly informative, but nonetheless need to be included in the identification strategy
to control for the fact that these populations were subject to a simultaneous treatment.
DD Estimates
Table 5 shows the β1 coefficient for each of the three AB 109 treatments. As mentioned pre-
viously, indicators for the timing of reception adjustment are always included as controls and X
represents a full set of controls including population shares and program participation. Columns
(3) and (4) are two different extensions of the specification in column (2). Column (3) adds time
fixed effects while column (4) adds a flexible time trend and omits the first three months of post-
treatment observations. In all specifications standard errors are clustered at the prison level and
each observation is weighted by the average population size of that prison measured over the six
29
Table 5: Difference-in-Differences Estimation
Dependent Variable: Log Rate of Assaults
(1) (2) (3) (4)VARIABLES Base Controls TimeFE 3mo.Gap
TreatLv2*Post -0.529*** -0.388** -0.390*** -0.446***(0.176) (0.142) (0.122) (0.159)
TreatLv3*Post -0.303 -0.211 -0.206 -0.306(0.279) (0.271) (0.144) (0.310)
TreatRec*Post 0.252 0.272 0.218 0.471***(0.191) (0.179) (0.173) (0.127)
Observations 1,470 1,470 1,470 1,380Controls RA only X X XTrend/Gap No No No Yes
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
months prior to AB 109 implementation.
The relationship between pre-treatment share of level 2 population and violence is consistently
negative and quite large. On the other hand, the coefficients for the other two treatments are either
not significant or even positive. Yet the positive coefficient on the reception treatment is consistent
with a large compositional effect driving an increase in assaults that overwhelms any decrease from
reduced crowding. Examining columns (2) and (4) reveals that omitting the few months of AB
109 prior to the reception adjustment increases both the magnitude and significance of the point
estimate for the reception treatment. This also aligns with the implications of the model in Section
4 since the omitted months in column (4) are the months for which crowding effect (decreasing
violence via EV ;c > 0) as well as the compositional effect (increasing violence via Eπ;λ < 0) can
both be expected to be quite large. By contrast, during the later months the reception adjustment
offsets much of the decreased crowding but the change in composition is sustained, making the
expected compositional effect stronger relative to the crowding effect.
The point estimates for the level 3 treatment are negative and qualitatively similar to the level
2 estimates but with large standard errors. Since the nature of the compositional bias in the level 3
treatment is unknown, it is difficult to gage the informative value of the resulting point estimates.
However, they do not contradict the evidence embodied by the coefficient for the level 2 treatment.
30
Absent the theory of compositional change, the estimates in Table 5 suggest that the decrease
in crowding due to AB 109 led to a decrease of approximately 40% in the rate of assaults at level 2
facilities. The level 2 facilities saw crowding fall about 30 percentage points from an initial point of
almost 200% of design capacity, implying a semi-elasticity of approximately 1.3 for this specific type
of prison population. On the other hand, there is no clear effect and possibly an increase in assaults
for the reception center populations. It is possible that this is due to some fundamental difference
between reception centers and level 2 facilities, either with regard to the populations themselves
or the housing and security protocols. Yet it is also true that the difference in estimates for the
level 2 and reception populations comport well with the assumption that there is compositional
element to the population shock generated by AB 109, because such compositional change would
be necessarily more significant among the reception population. The results can thus be interpreted
as descriptive evidence supporting the presence of such a compositional effect.
6.2 Instrumental Variable (IV)
Cit = α0Monthst + α1Months2t + α2Monthst ∗ Sni + α3Xit + α4f(t) + γi + uit (3)
Yit = β0 + β1Cit + β2Xit + β3f(t) + δi + εit (4)
Equations 3 and 4 present the basic structural form of the IV strategy. Equation 3 is the first
stage estimating equation wherein crowding is estimated as a function of the number of months since
the implementation of the policy, Monthst, interacted with the pre-implementation shares, Sni , of
population type n. In the baseline IV, only initial population shares for the treated subpopulations
are used as instruments. Specifications tested with an expanded set of instruments find only minor
adjustments to the coefficient of interest. Equation 4 is the second stage estimation, which is a
fixed effects regression of the rate of assaults on the predicted values of the crowding variable.
The estimation includes the same control variables included in the main DD specification20. A
polynomial time trend is included in each stage of estimation.
In effect, the IV strategy uses the time since implementation of AB 109 and the initial share of
20These include indicators for the timing of the reception adjustment and the population share and programenrollment variables included in Table 1.
31
each population type to predict the change in crowding at each prison21. This approach exploits
the same exogenous variation as the DD approach but allows flexibility in modeling variation in
the intensity of treatment over time. The IV strategy also has the benefit of added flexibility in
modeling the impact of the reception adjustment. Specifically, the “Interact RA” specification
in Table 6 allows the degree to which pre-shock reception share predicts changes in crowding to
be diminished as the reception adjustment occurs in the given prison. The major shortcoming of
the IV specification is that it necessarily conflates any compositional bias from any of the three
treatments into the estimated effect of crowding on the rate of assaults.
The exclusion restriction for this IV strategy requires that the pre-shock composition of prison
populations is uncorrelated with future variation in the rate of assault and battery, other than
through it’s correlation with changes in the level of crowding due to the policy design. The only clear
challenge to the validity of this restriction is the afore mentioned fact that pre-AB 109 population
shares are correlated with the post-AB 109 changes in those shares. For this reason all specifications
of the IV model include a full set of controls than include contemporaneous population shares for
each subpopulation.
IV Estimates
Table 6 provides the estimated β1 for a number of variations on the IV model. A benefit of the
IV model is the straightforward interpretation of the β1 coefficient. It is a semi-elasticity showing
the marginal effect of crowding on the rate of assaults, where crowding is measured in percentage
point changes and the rate of assault in percent changes. The first column of the table is the base
specification of the model, exactly as presented in equations 3 and 4. The other three columns each
represent a different variation from the base specification. The second column includes an additional
term in the excluded instruments, which interacts the RAit variable22 with the Monthst ∗Sni term
of the reception treatment. Column (3) is a robustness check that removes the Months terms
that are not interacted with population shares from the IV exclusions, allowing that there may be
some statewide correlation between the policy timing and assaultive behavior. The standard error
inflates slightly with this change, but the coefficient remains large and statistically significant. In
21These population shares are defined as the population of type n divided by the total prison population, averagedover the 6 months preceding implementation of AB 109.
22RAit is an indicator that turns on in the month that prison i has the reception adjustment.
32
Table 6: IV Model with Prison Fixed Effects
Dep. Variable: Log of Assaults and Batteries per 100 inmates
(1) (2) (3) (4)VARIABLES Base IV Interact RA Exclusion Drop 3mo.
Crowding (P/K) 2.117*** 1.548*** 1.718** 2.289***(0.627) (0.497) (0.767) (0.629)
Observations 1,440 1,440 1,440 1,350Number of ID 30 30 30 30Controls X X X+Months XExclusions Basic Months*S*RA Months*S BaseF test IVs 10.34 26.81 10.05 10.59
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
the final column the 3 month period just following implementation and preceding full scale reception
adjustment is excluded from the analysis. This is an alternate means of accounting for the effects of
the reception adjustment, so the RA control variable is excluded from the specification in column
(4).
A few things are immediately observable from these results. First, a comparison of these es-
timates with the earlier OLS estimates shows evidence of significant downward bias in the OLS
estimates, likely due to endogeneity as discussed earlier. The estimates from each of the IV spec-
ifications are large, statistically significant, and reasonably stable. These estimates are possibly
subject to significant attenuation due to compositional bias, yet still exhibit semi-elasticities as
high 2.2. The more conservative estimate in column (2) suggests that decreasing crowding by 10
percentage points can decrease the rate of assault by as much as 15 percent.
The difference between the column (1) and (2) estimates is also of interest. The two estimations
are identical beyond the one addition of the interactionMonthst∗SRi ∗RAit. As discussed previously,
this interaction provides a greater ability to predict changes in crowding using the pre-shock share
of reception population. In the absence of this interaction, the controls for administrative response
only allow for a level shift in crowding once there has been a response. In the first stage regression
coefficients (available in Table 10 in the appendix), the point estimate for the coefficient on this
interaction is almost precisely the same magnitude as the α2 for reception share, which is the
33
negative of the former. This implies that the pre-shock share of reception population correctly
predicts decreases in crowding when AB 109 begins and up until the reception adjustment occurs,
at which point it ceases to have any predictive power at all because the two coefficients cancel
each other out. This improved fit in the first stage modeling of crowding via the reception share
together with the diminished coefficient in Table 6, aligns with the results from the DD strategy
that reception centers do not see assaults decrease to the degree, if at all, that the other treated
facilities do.
6.3 Tests & Robustness
Section 2 presents the argument that previous research has struggled to identify a consistent corre-
lation between prison crowding and violent behavior. Further, because of the correlational nature
of much of the existing research, the inconsistency between studies implies that even when a pos-
itive correlation is identified it is as likely due to spurious correlation or publication bias as any
true relationship. This research contributes to the literature estimates driven by AB 109, a court
mandated policy shock that directly impacted the level of crowding in California prisons; AB 109
provides rationale for the position that the estimates in this paper are unbiased evidence of the
causal effect of crowding on violent behavior. Yet a plausible source of bias remains, namely the
compositional effect defined in Section 4. However, the compositional effect is reasonably expected
to bias estimates towards zero and thus at worst the estimates in this paper would be a lower bound
on the true value. Still the limitations of these data and the exogeneity of the AB 109 intervention
do warrant some further discussion.
The first concern pertains to the validity of the channel by which the instruments and DD design
identify the effect of crowding on violence. Although it is certainly the true that AB 109 reduced
crowding, it remains a possibility that there is some spurious artifact of the assault data driving the
results or some unknown features to the implementation of the new law that were the real causes
of reduced violence. For example, the inability to control for staffing changes, due to limitations
of these data, is a potential concern. Furthermore, even if staffing changes were insignificant23, it
is possible that staff were simply able to operate more effectively once there were fewer inmates
23Although the available data on staffing has missing data and other issues that make the author uncomfortableincluding staffing variables in this research, a general read of the existing data suggests that staffing changes due toAB 109 were minor.
34
Table 7: IV Model: Testing other measures of misconduct.
Dep. Variable: Log-rate of the given form of misconduct.
(1) (2) (3) (4)VARIABLES Assaults Incidents Drugs Cellphone
Crowding (P/K) 1.522*** 1.406*** 0.418 -0.0583(0.558) (0.312) (0.761) (0.810)
Observations 750 750 750 750Number of ID 30 30 30 30Controls X X X XExclusions Basic Basic Basic BasicF test IVs 16.91 16.91 16.91 16.91
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
crowding the facilities. Yet if the policy induced variation in assaults were due to enforcement gains
from improved staffing ratios or effectiveness of staff then the estimated effects should be present
for other forms of misconduct as well.
Table 7 confronts the above concerns by repeating the IV specification with different measures
of misconduct. Some of these measures were not recorded in the early years of CompStat data,
so the number of pre-AB 109 observations in this table are limited to 15 months. Column (1)
repeats the IV specification from column (2) of Table 6 for the limited time period of the other
specifications here, showing a qualitatively identical estimate to those in Table 6. Column (2)
replaces the measure of assaults, previously disciplinaries, with the “incidents” measure of the
same type of violation. The point estimate is not sensitive to the which measure of assaults is
used. In stark contrast, columns (3) and (4) prove there is no identifiable impact on other forms
of misconduct that occur with similarly high frequency to that of assaults.
The validity of using the variation in crowding from AB 109 is further tested in Table 8. Column
(1) is a replication of column (1) from Table 6 and each of the following columns repeats the model
with the policy implementation beginning the specified number of months earlier than the true
implementation date of AB 109. To maintain comparability of the estimates, the data is truncated
in each specification so the number of post-implementation months remains constant. Since there
was no equivalent in these periods to the reception adjustment that occurred following AB 109,
35
Table 8: IV Model: Placebo tests with policy implementation at alternate dates.
Dep. Variable: Log of Assaults and Batteries per 100 inmates
(1) (2) (3) (4) (5)VARIABLES Base IV 9 Month 15 Month 21 Month 27 Month
Crowding (P/K) 2.117*** 1.891 0.791 0.356 0.195(0.627) (1.998) (1.097) (0.901) (2.196)
Observations 1,440 1,140 960 780 600Number of ID 30 30 30 30 30Controls X X X X XExclusions Basic Basic Basic Basic BasicF test IVs 10.34 2.874 5.248 2.761 2.054
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
the RA variable is omitted from the lagged specifications. The results are no different if the RA
control is reintroduced into the estimating equations. None of the resulting estimates are close to
statistical significance in either the first or second stage regressions.
Tables 7 and 8 both rely on the IV strategy to test their respective alternate hypotheses. This
approach is more direct and has the benefit of reporting fewer coefficients, whereas the multiple
coefficients reported in the DD strategy increases the risk of spurious statistical significance. In
addition, the less refined modeling of time variation in the DD strategy is more likely to pick up
correlations from any of several earlier, less dramatic policy changes that occurred in the years
leading up to AB 109.
It is also possible that while overall variation in crowding from AB 109 is appropriately exoge-
nous, the subpopulations used to identify the intensity of crowding reductions are not the proper
channel. The likely alternative is that prison administrators were able to manipulate new prisoner
classifications to reduce populations in facilities that were the most crowded, rather than those
naturally targeted by reducing the inflow of non-violent offenders. The result of this would be that
initial crowding should do a better job predicting the post-AB 109 reductions in crowding than the
pre-shock population shares do.
Figures 7 and 8 depict the relationship between the initial level of crowding and change in
crowding, with Figure 8 breaking the trends up by the same rough treatment groups used in Table
36
-80-60
-40-20
0
Chan
ge in
Perc
ent B
eyon
d Cap
acity
40 60 80 100 120 140Percent Beyond Design Capacity
Post-AB 109 Change in Crowding by Initial Crowding
Figure 7: This figure shows the correlations between initial crowding and the decrease in crowding afterimplementation of AB 109.
-80-60
-40-20
0
Chan
ge in
Perc
ent B
eyon
d Cap
acity
40 60 80 100 120 140Percent Beyond Design Capacity
Control Group Level 2 Group Level 3 Group Reception Group
Post-AB 109 Change in Crowding by Initial Crowding
Figure 8: This figure shows the correlations between initial crowding and the decrease in crowding afterimplementation of AB 109, separated by “treatment groups”. It shows that the apparent correlation betweenthe variables in Figure 7 is due to the fact that the treated populations were, on average, those that weremore crowded to begin with.
37
-.6-.4
-.20
.2C
hang
e in
the
Rat
e of
Ass
aults
-80 -60 -40 -20 0Change in Percent Beyond Capacity
Control Group Level 2 Group Level 3 Group Reception Group
Post-AB 109 Change in Assaults by Change in Crowding
Figure 9: This figure shows the correlations between change in crowding and the change in the rate ofassault following implementation of AB 109, separated by “treatment groups”.
38
4 and Figure 6. If the above concern were valid, then the correlation apparent in Figure 7 would
remain mostly intact in Figure 8, with the treated populations simply having systematically higher
levels of initial crowding (which would then induce the larger reductions seen for those groups
in Table 4). Instead the fitted lines for each group in Figure 8 have relatively shallow gradients
and those observations in the treatment groups with less crowding saw larger reductions than
comparable prisons in the untreated group. The exception to the shallow slopes in Figure 8 is the
line for the reception group, for which the fitted line retains a slope similar to that in Figure 7.
However, note that the slope for the reception group is driven by the single outlier in the bottom
right of the figure24. With the outlier removed, the reception line has a slope as shallow as any
of the others. In sum, when treatment groups are accounted for there is still a weak correlation
between initial crowding the the reduction induced by AB 109, but the treated subpopulations
appear to be an appropriate predictor of changes in crowding.
A final issue for discussion is the omitted outlier from Figure 6. The omission has a significant
impact on the presentation of that figure, which is shown by Figure 9. The outlier, DVI, experiences
the largest reduction in crowding among the prisons but also experiences the largest increase in
assaults, which dramatically alters the fitted line for the reception group. A brief investigation of
this did not reveal a clear reason for the distinct experience of DVI relative to the other prisons.
As such, DVI was not excluded from any of the empirical specifications in this section, only the
earlier figure to better illuminate the consistent relationship in the data. However, it is also of note
that exclusion of DVI does not qualitatively change any of the empirical estimates. Table 11, in
the appendix, demonstrates this by replicating the full set of DD specifications with DVI excluded
from the analysis.
7 Conclusion
Violent behavior in prisons may be caused by a variety of mechanisms. In line with much of the
literature (Kearney et al., 2014), this paper examines the role of prison crowding as a determinant
of violence. This relationship plays an important role in well-informed policy decisions and its
24The outlier in this figure is again DVI, the same prison that was omitted from Figure 6
39
study is particularly apropos given recent political movements to reduce mass incarceration in the
United States.
Although theory and popular opinion have long held prison crowding as a key determinant of
violent behavior, empirical estimates in the existing literature have been surprisingly inconsistent
(Franklin, Franklin, & Pratt, 2006). The estimates in this paper are the first to use a quasi-
experimental design that specifically targets the effect of crowding on violence. These estimates
represent persuasive evidence of a causal effect of crowding on violent misconduct that is positive
and qualitatively large. They suggest a 10 percentage point decrease in prison crowding results
in a decrease in the rate of violent assaults by approximately 15%, which could imply significant
cost savings associated with reductions in crowding. A careful review of the empirical results
also provides evidence that the compositional effect defined in Section 4 is indeed a factor in the
outcomes from the AB 109 shock to crowding. This further implies that the compositional effect is a
plausible factor to be considered with regard to previous and future empirical work on crowding and
violence. The compositional effect, as a new source of potential bias, adds a nuanced explanation
for conflicting evidence in existing empirical research on crowding and violence.
There are some natural limits to the implications of this research in policy application. Foremost
among these is external validity when considering dissimilar prison populations. Although the IV
strategy incorporates some statewide variation in crowding across prisons, the estimates are largely
driven by variation in security level 2 populations. These populations have relatively low rates of
violent misconduct and it is possible that prison populations with a higher propensity for violence
could be either more or less responsive to changes in crowding. The DD estimates for the other
treated populations do little to better inform this issue since both are expected to be subject to
greater compositional bias than the level 2 treatment. It is also important to recognize that the
identifying variation for this research is from very high levels of overcrowding, many of the affected
facilities beginning well above 200% capacity prior to implementation of AB 10925. It is reasonable
to expect some variation in the marginal impact of crowding on violent behavior when examining
the relationship in much less crowded facilities.
In summary, this research offers several unique contributions to the literature. It presents the
first empirical evidence of a strong causal effect of prison crowding on violent misconduct. It
25The initial levels of crowding can be observed along the horizontal axis in Figures 7 and 8.
40
also introduces the idea that there is a compositional element to policy interventions that reduce
crowding, which can lead to bias in empirical estimates. Although this concept is not necessarily
novel, having come up in many of the author’s conversations on the topic, this paper is the first to
formally discuss the idea and its role in studying prison crowding and violence. Finally, this work
highlights many areas of opportunity for subsequent work studying prison violence.
One issue to be further addressed in future research is heterogeneous effects among different
types of prison populations. As mentioned to earlier, the estimates in this paper are mostly per-
tinent to prison populations that are relatively less prone to violence and prison settings that are
subject to extreme levels of crowding. Maximum security facilities are of particular interest in this
regard, both because they have the highest baseline rates of violence26 and because housing and
security protocols in such facilities tend to be quite different from other prison facilities.
Another important extension of this research will be to access improved data to allow an in
depth analysis of the interaction of the compositional and crowding effects from changes in prison
populations. With sufficiently detailed individual inmate data, a full decomposition of the respective
mechanisms is possible and will allow a much more nuanced understanding of policy impacts on
violent misconduct. This information could be an invaluable contribution to social, political, and
administrative insights regarding the costs and benefits of a broad variety of justice related policy
reforms.
26Security level 4 facilities in California have significantly higher rates of violence than other facilities, show inTable 9.
41
References
Adams, K. (1983). Former mental patients in a prison and parole system; a study of socially
disruptive behavior, Criminal Justice and Behavior 10(3): 358–384.
Blevins, K. R., Listwan, S. J., Cullen, F. T. and Jonson, C. L. (2010). A general strain theory of
prison violence and misconduct: An integrated model of inmate behavior, Journal of Contempo-
rary Criminal Justice 26(2): 148–166.
Boylan, R. T. and Mocan, N. (2014). Intended and unintended consequences of prison reform,
Journal of Law, Economics, and Organization 30(3): 558–586.
Camp, S. D. and Gaes, G. G. (2005). Criminogenic effects of the prison environment on inmate
behavior: Some experimental evidence, Crime & Delinquency 51(3): 425–442.
Camp, S. D., Gaes, G. G., Langan, N. P. and Saylor, W. G. (2003). The influence of prisons on
inmate misconduct: A multilevel investigation, Justice Quarterly 20(3): 501–533.
Cao, L., Zhao, J. and Van Dine, S. (1997). Prison disciplinary tickets: A test of the deprivation
and importation models, Journal of Criminal Justice 25(2): 103–113.
Carroll, L. (1974). Hacks, blacks, and cons: Race relations in a maximum security prison, Lexington
Books Lexington, MA.
Caudill, J. W., Trulson, C. R., Marquart, J. W., Patten, R., Thomas, M. O. and Anderson, S.
(2014). Correctional destabilization and jail violence: The consequences of prison depopulation
legislation, Journal of Criminal Justice 42(6): 500–506.
Chen, M. K. and Shapiro, J. M. (2007). Do harsher prison conditions reduce recidivism? a
discontinuity-based approach, American Law and Economics Review 9(1): 1–29.
Cochran, J. C. (2012). The ties that bind or the ties that break: Examining the relationship
between visitation and prisoner misconduct, Journal of Criminal Justice 40(5): 433–440.
Conrad, J. P. (1966). Violence in prison, The ANNALS of the American Academy of Political and
Social Science 364(1): 113–119.
42
DeLisi, M., Berg, M. T. and Hochstetler, A. (2004). Gang members, career criminals and prison
violence: Further specification of the importation model of inmate behavior, Criminal Justice
Studies 17(4): 369–383.
Ellis, D., Grasmick, H. G. and Gilman, B. (1974). Violence in prisons: A sociological analysis,
American Journal of Sociology pp. 16–43.
Flanagan, T. J. (1983). Correlates of institutional misconduct among state prisoners, Criminology
21(1): 29–40.
Franklin, T. W., Franklin, C. A. and Pratt, T. C. (2006). Examining the empirical relationship
between prison crowding and inmate misconduct: A meta-analysis of conflicting research results,
Journal of Criminal Justice 34(4): 401–412.
Gaes, G. G. and McGuire, W. J. (1985). Prison violence: The contribution of crowding versus other
determinants of prison assault rates, Journal of research in crime and delinquency 22(1): 41–65.
Gaes, G. G., Wallace, S., Gilman, E., Klein-Saffran, J. and Suppa, S. (2002). The influence of prison
gang affiliation on violence and other prison misconduct, The Prison Journal 82(3): 359–385.
Griffin, M. L. and Hepburn, J. R. (2006). The effect of gang affiliation on violent misconduct among
inmates during the early years of confinement, Criminal Justice and Behavior 33(4): 419–466.
Huebner, B. M. (2003). Administrative determinants of inmate violence: A multilevel analysis,
Journal of Criminal Justice 31(2): 107–117.
Irwin, J. and Cressey, D. R. (1962). Thieves, convicts and the inmate culture, Soc. Probs. 10: 142.
Kearney, M. S., Harris, B. H., Jacome, E. and Parker, L. (2014). Ten economic facts about crime
and incarceration in the united states, The Hamilton Project, May. Accessed November 4: 2014.
Lahm, K. F. (2008). Inmate-on-inmate assault a multilevel examination of prison violence, Criminal
Justice and Behavior 35(1): 120–137.
MacKenzie, D. L. (1987). Age and adjustment to prison interactions with attitudes and anxiety,
Criminal Justice and Behavior 14(4): 427–447.
43
McCorkle, R. C., Miethe, T. D. and Drass, K. A. (1995). The roots of prison violence: A test
of the deprivation, management, and “not-so-total” institution models, Crime & Delinquency
41(3): 317–331.
Megargee, E. I. (1977). The association of population density, reduced space, and uncomfortable
temperatures with misconduct in a prison community, American journal of community psychology
5(3): 289–298.
Myers, L. B. and Levy, G. W. (1978). Description and prediction of the intractable inmate, Journal
of Research in Crime and Delinquency 15(2): 214–228.
Orrick, E. A. and Morris, R. G. (2015). Do parole technical violators pose a safety threat? an
analysis of prison misconduct, Crime & Delinquency 61(8): 1027–1050.
Petersilia, J., Honig, P. and Hubay, C. (1980). The prison experience of career criminals, US
Department of Justice, National Institute of Justice.
Petersilia, J. and Snyder, J. G. (2013). Looking past the hype: 10 questions everyone should ask
about california’s prison realignment, California Journal of Politics and Policy 5(2): 266–306.
Stephan, J. J. (1999). State prison expenditures, US Department of Justice, Office of Justice
Programs, Bureau of Justice Statistics.
Sykes Gresham, M. (1958). The society of captives. a study of a maximum security prison.
Wheeler, S. (1961). Socialization in correctional communities, American Sociological Review
pp. 697–712.
White, A. A. (2008). Concept of less eligibility and the social function of prison violence in class
societies, the, Buff. L. Rev. 56: 737.
Wooldredge, J. D. (1991). Correlates of deviant behavior among inmates of us correctional facilities,
Journal of Crime and Justice 14(1): 1–25.
Wooldredge, J., Griffin, T. and Pratt, T. (2001). Considering hierarchical models for research on
inmate behavior: Predicting misconduct with multilevel data, Justice Quarterly 18(1): 203–231.
44
Wright, K. N. (1989). Race and economic marginality in explaining prison adjustment, Journal of
Research in Crime and Delinquency 26(1): 67–89.
Wright, K. N. (1991). The violent and victimized in the male prison, Journal of Offender Rehabil-
itation 16(3-4): 1–26.
45
A Theory Appendix
Return to Section 4.
The objective of this model is to capture the major channels by which policy interventions that
increase or decrease the population (and thus crowding) may impact violent behavior in prisons.
The main channel of interest is the role of changing the composition of the prison population with
regards to individuals’ tendency to resort to violence. To this end, it is not the intent of this model
to explore the motives underlying individual choices. Nor is it intended that the model capture
the determinants of all forms of violent behavior. Since premeditated assaults and strategic gang
violence are unlikely to be responsive to the crowdedness of the prison, the model is oriented to
violent behavior may be seen as impulsive or extemporaneous.
Recall that the three mechanisms defined in Section 4 were a behavioral mechanism, a structural
mechanism, and a compositional mechanism. The setting is distilled into a reduced form, applied
probability model where pairwise interactions occur between inmates. These interactions are as-
sumed to be contentious in nature and each has a probability of resulting in violence. Since the
interest of this model is the effect of changing the composition of a set of heterogeneous individuals,
the underlying rational choice problem of each individual is suppressed in this presentation of the
model27. Instead, each individual is assumed to have a baseline propensity for violence, ai, that is
distinct from their probability of resorting to violence in any given situation. Propensity is assumed
to be a fixed individual characteristic, whereas the probability is a function of situational factors
(e.g. crowding) and the individual’s fixed propensity. However, it is the overall rate of violence
for the prison population that is modeled in this appendix, so the implications of these individual
propensities are aggregated to the prison level variables in the following model set up.
V Total violence per unit time.
P Total prison population.
K Design capacity of the prison.
c Degree of crowding, defined as the population/capacity ratio (P/K).
s Scale effect as a function of P , generally taken to exhibit constant returns to scale.
27Future work on this research will include treatment of the underlying choice setting for inmates, as well as thestrategic interaction between inmates that leads to taunting and baiting their opponents. As it pertains to the broaderidea of compositional change, the author does not believe those stages to be instrumental.
46
n Number of potentially violent interactions per unit time for each individual, with n′(c) ≥ 0.
π Probability that a particular interaction will be violent, with π′(c) ≥ 0.
λn A shift parameter for policy ‘n’ with a positive relationship to crowding.
ai An index measuring individual i’s baseline propensity for violence.
F (·) The distribution of all inmates’ ai, with support [0, 1]28.
Given this setting, the aggregate incidence of violence in the prison is defined by the identity in
Equation 5. The equation simply states that the total violence in a unit of time will equal the total
number of pairwise interactions that occur per unit time29 multiplied by the average probability
that a single interaction becomes violent. This very basic representation of the setting allows
the structural crowding mechanism to be captured by n(c). Meanwhile π(c, λn) captures both the
behavioral and compositional mechanisms, respectively the direct effect of c and the marginal effect
of the shift parameter λn.
V = [s(P )][n(c)][π(c, λn)] (5)
To derive the responsiveness of violence to crowding and incorporate the principle that changes
in crowding are the result of some policy change, it is assumed that crowding (and thus total popu-
lation) is a function of the policy parameter (λn) with an elasticity of one (Ec:λn = 1). Furthermore,
we allow that the relationship between the shift parameter (λn) and probability of violence (π(·))
can vary between policy options. That is, the partial derivatives ∂π∂λ1
< 0 and ∂π∂λ2
> 0 are explicitly
permitted in this setting.
The elasticity of violence with respect to a population shock is derived as follows.
V = s(P )n(c)π(c, λn)
= s(cK)n(c)π(c, λn)
=⇒ ln[V ] = ln[s(cK)] + ln[n(c)] + ln[π(c, λn)]
=⇒ 1
V
dV
dλn=
1
s(P )s′(P ) ·K dc
dλn+
1
n(c)n′(c)
dc
dλn+
1
π(c, λn)
[∂π∂c
dc
dλn+
∂π
∂λn
]28F (·) is an approximation of the true distribution from which an inmate’s “partner” in any pairwise interaction
is randomly is randomly drawn. For sufficiently large populations, the distributions are effectively identical.29Under CRS, s(P ) · n(c) can be simplified to just P · n(c), which is the total population times the number of
interactions per individual per unit time.
47
Then multiply through by λn.
λn
V
dV
dλn=[ 1
s(P )s′(P ) ·K +
1
n(c)n′(c) +
1
π(c, λn)
∂π
∂c
] dcdλn
λn
cc+
λn
π(c, λn)
∂π
∂λn
=⇒ EV :λn =[ 1
s(P )s′(P ) ·K +
1
n(c)n′(c) +
1
π(c, λn)
∂π
∂c
]c · Ec:λn + Eπ:λn
=[ cKs(P )
s′(P ) +c
n(c)n′(c) +
c
π(c, λn)
∂π
∂c
]Ec:λn + Eπ:λn
=[Es:P + En:c + Eπ:c
]Ec:λn + Eπ:λn
Now recall that λn is defined such that Ec:λn = 1 and note that constant returns to scale30
requires that Es:P = 1. Then the total policy impact can be decomposed into the scale effect plus
an elasticity that represents each of the three mechanisms, as shown in Equation 6.
EV :λn = 1 + En:c + Eπ:c + Eπ:λn (6)
The first terms on the righthand side of Equation 6 constitute the direct elasticity of violence
with respect to crowding, EV :c. Therefore Equation 7, the same as presented in Section 4, is an
equivalent representation of the relationship between violence and the policy parameter.
EV :λn = EV :c + Eπ:λn (7)
There is much yet to be discovered about the motivations and determinants of violent behavior,
in and out of prisons. What is offered here is the simple insight that there are several potential
mechanisms at work that determine the net effect of changing prison crowding on violent behavior.
Furthermore, with sufficiently rich data, it should be possible to perform a decomposition of the
overall impact of a policy intervention and differentiate between these mechanisms. Lastly, in
the absence of very granular data, estimates using any significant variation in crowding should be
viewed as estimates of the relationship represented by EV :λn and not EV :c.
There is an argument to be made that the latter point is not necessarily an issue. For example,
in a study that is a straightforward impact evaluation following a new law or regulation, the effects
of each mechanism are all part of the impact that the law had on violence and thus rightfully
30CRS in this setting implies that, in the absence of any crowding or compositional changes, doubling the populationsize will double the total amount of violence.
48
included in the analysis. The caution raised by this model is in proper interpretation. AB 109 and
the results of this research provide a poignant example of this. Separate impact evaluations for
reception centers and level 2 facilities would likely show no significant impact on violence in the
former and a decrease in the rate of violence for the latter. Attributing these directly to crowding
would imply that violent behavior in reception populations is not responsive to crowding. While
the model presented here does not preclude that possibility, it does raise a plausible alternate
explanation that simultaneously accounts for the difference in outcomes at level 2 facilities.
The framework is also helpful for conceptualizing the role that the design of an intervention
plays. The impact of AB 109 on reception centers is a good case of this. In the initial stage, AB 109
dramatically reduces crowding in reception facilities and does so by eliminating the flow of incoming
non-violent offenders. Since reception centers take custody of all other offenders upon arrival in
the prison system, now mostly very serious offenders, it can be expected that there is a significant
compositional change in addition to the large decrease in crowding. These are expressed in Equation
6 as En:c > 0 and/or Eπ:c > 0 from the crowding mechanisms and Eπ:λn < 0 from compositional
change. But there is also the reception adjustment, which consolidates the remaining reception
populations into fewer facilities. This diminishes the overall drop in crowding, which pertains to
the first two elasticities, but maintains the compositional change associated with the very large
reduction in total reception population size, which acts through the latter elasticity.
In summary, empirical work on the relationship between prison crowding and violence typically
claims to estimate that relationship directly. The model presented here questions whether that
is really the case, based on two assumptions: there is hetergeneity among inmates with regards
to their propensity for violence and increasing or decreasing crowding is associated with altering
the composition of the inmate population with respect to this heterogeneity. Where possible a
decomposition of the policy effects on violence should be performed to properly illuminate the
distinct roles of crowding and composition. With less precise data, logic dictates that some policies
will have a predictable pattern of selection by which they cause compositional change and this can
be used to make inferences about the manner in which compositional change may bias estimates
from the data.
Return to Section 4.
49
B Empirical Appendix
This appendix provides figures and regression estimates that have been excluded from the main
body of the paper. This additional information tests alternate specifications of the models and
provides additional context to the California prison setting.
An interesting piece of context for understanding violence in California prisons is the difference
in baseline rates of violence between the subpopulations. Due to the same data limitations with
units of observation that constrain the main identification31, exact averages by subpopulation are
not possible. Instead, Table 9 reports a basic OLS regression of the rate of assault on shares
of each major subpopulation. The regression is estimated without a constant term, so the point
estimates can be viewed as rough approximations of the average rate of assault for facilities with
that subpopulation. The negative coefficient on the special needs population emphasizes the fact
that there are confounding factors in these approximations. Nonetheless, the estimates follow an
intuitive pattern that should be expected. The rates of violence are monotonically increasing in
security classification and reception centers fall within bounds set by the security levels. Even high
rate at reception centers relative to level 2 and 3 facilities aligns well with the idea that stability
in reception centers is disrupted by the high rate of turnover.
Table 10 presents the first stage estimates of the IV specifications from Section 6. The level 2
and level 3 instruments are very consistent predictors of decreased crowding across all specifications.
The interesting variation comes in the reception instrument (Months ∗Rec) and the months since
implementation variable (Months). Adding the additional RA interaction with the reception in-
strument in column (2) dramatically increases the significance and the magnitude of the coefficient
on Months ∗ Rec. At the same time, column (2) is the only specification for which Months has
little correlation with decreased crowding. Furthermore, the coefficients on the two terms for the
reception instrument are the inverse of each other, neatly offsetting one another once a reception
adjustment occurs. Together this suggests that the pre-AB 109 reception share is correlated with
large reductions in crowding during the early months of the new law, but this correlation dissipates
as reception adjustments begin to offset the crowding reductions. When the specifications do not
allow for this mid-shock change in the marginal effect of the reception instrument, the correla-
31The unit of observation in the data is a prison. Each prison has several facilities that typically house differenttypes of subpopulation, therefore the rate of violence for each prison cannot be attributed to a single subpopulation.
50
Table 9: OLS Regression of Population Shares on Assaults
Dependent Variable: Rate of Assault per 100 Inmates
(1)VARIABLES Trend
Security Level 1 0.244***(0.0273)
Security Level 2 0.311***(0.0179)
Security Level 3 0.623***(0.0266)
Security Level 4 1.148***(0.0370)
Reception Center 0.934***(0.0236)
Special Needs -0.295***(0.0412)
Observations 1,440
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
tion with initial reductions in crowding due to reception share are spread between the reception
instrument and the Months variable which has a more consistent negative correlation.
Tables 11 and 12 are additional tests of the robustness of the DD estimation strategy. Table 11
simply replicates the Table 5 specifications, but excludes the outlier that is shown in Figure 9 and
excluded from Figure 6. The coefficients for the reception treatment are clearly not dependent upon
this outlier. Table 12 repeats the DD specification time fixed effects for alternate outcome variables,
as Table 7 did with IV strategy. Using incidents rather than disciplinaries to measure assaults
undermines the estimates for this strategy. As alluded to earlier, there remain unanswered questions
about the selection process by which incidents are reported because they are far less frequent than
disciplinaries. The significant coefficient for the reception treatment on drug possession is likely a
function of the law disrupting the channels by which contraband is funneled into prisons32.
Tables 13 and 14 test the effects on the IV strategy of changing the measure of violence. Table
32An anecdotal example of this: during an informal interview with a inmate in a California prison, the author wastold that one way prison gangs funnel drugs into the prison was to have someone out on parole swallow baggies, thenviolate their parole to bring them inside the prison. By redirecting parole violators to county jails, AB 109 disruptedthe part of the supply chain.
51
Table 10: 1st stage regressions from IV estimates in Table 6
(1) (2) (3) (4)VARIABLES Base IV Interact RA Exclusion Drop 3mo.
Months -0.222** -0.150 -0.222** -0.194*(0.0958) (0.0883) (0.0958) (0.1000)
Months sq 0.146* 0.0630 0.146* 0.132(0.0793) (0.0702) (0.0793) (0.0853)
Months*Lv2 -0.290*** -0.309*** -0.290*** -0.302***(0.0751) (0.0761) (0.0751) (0.0760)
Months*Lv3 -0.324*** -0.352*** -0.324*** -0.329***(0.0743) (0.0719) (0.0743) (0.0729)
Months*Rec -0.158* -1.223*** -0.158* -0.0733(0.0929) (0.195) (0.0929) (0.0883)
Months*Rec*RA 1.210***(0.216)
Observations 1,440 1,440 1,440 1,350Number of ID 30 30 30 30Controls X X X X
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Table 11: DD Estimation: DVI (outlier) excluded from observations.
Dependent Variable: Log Rate of Assaults
(1) (2) (3) (4)VARIABLES Base Controls TimeFE 3mo.Gap
TreatLv2*Post -0.537*** -0.405*** -0.406*** -0.459***(0.176) (0.144) (0.123) (0.165)
TreatLv3*Post -0.304 -0.217 -0.211 -0.307(0.278) (0.267) (0.145) (0.307)
TreatRec*Post 0.241 0.277 0.207 0.398***(0.187) (0.176) (0.180) (0.0993)
Observations 1,421 1,421 1,421 1,334Controls None X X XTrend/Gap No No No Yes
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
52
Table 12: DD Model Placebo Tests
Dep. Variable: Log Rate of the given form of misconduct.
(1) (2) (3) (4)VARIABLES Assaults Incidents Drugs Cellphone
TreatLv2*Post -0.314** -0.175 -0.162 -0.210(0.156) (0.143) (0.178) (0.242)
TreatLv3*Post -0.233 -0.0917 0.00244 0.197(0.175) (0.160) (0.200) (0.273)
TreatRec*Post 0.0585 -0.0689 -0.550** -0.0891(0.204) (0.187) (0.233) (0.320)
Observations 780 780 780 779Controls X X X XTrend/Gap No No No No
Standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Table 13: IV Model: Murder and attempted murder included in outcome variable
Dep. Variable: Log Rate of Violence per 100 inmates
(1) (2) (3) (4)VARIABLES Base IV Interact RA Exclusion Drop 3mo.
Crowding (P/K) 2.208*** 1.656*** 1.911** 2.338***(0.604) (0.481) (0.755) (0.605)
Observations 1,440 1,440 1,440 1,350Number of ID 30 30 30 30Controls X X X+Months XExclusions Basic Months*S*RA Months*S BaseF test IVs 10.34 26.81 10.05 10.59
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
53
Table 14: IV Model: Staff assaults excluded from outcome variable
Dep. Variable: Log Rate of Inmate-on-inmate Assault
(1) (2) (3) (4)VARIABLES Base IV Interact RA Exclusion Drop 3mo.
Crowding (P/K) 1.763** 1.123* 1.521* 1.958***(0.752) (0.605) (0.872) (0.744)
Observations 1,440 1,440 1,440 1,350Number of ID 30 30 30 30Controls X X X+Months XExclusions Basic Months*S*RA Months*S BaseF test IVs 10.34 26.81 10.05 10.59
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
13 includes murder and attempted murder with the original measures of assault and battery. This
leads to slightly larger point estimates, but no substantive changes in outcomes. Table 14 excludes
assaults on staff members from the outcome variable and finds diminished statistical significance
and point estimates in each specification. This suggests that variation in the rate of aggression
towards staff is a dimension in which crowding impacts violence.
The effects of the reception adjustment on alternate subpopulations is illustrated in Figures 10
through 13. Although there is some time variation following AB 109 for each of these, none of them
follow the distinct pattern shown for the security level 3 population.
Figure 14 shows the change in several measures of misconduct, measured over the 6 months
preceding implementation of AB 109 and the 4th through 9th months following it. This table
provides context for the decision to focus on misconduct measured via rates of assault. The other
forms of misconduct are generally less stable and subject to more uncertain sources of variation.
Figure 15 shows changes in the participation rate (per 100 inmates) in the programs included in
the standard set of controls. For these some increases are expected since population is decreasing
and there is no reason to expect a decrease in program capacities. The two most notable things in
this figure are the decrease in academic enrollment and the large increase in SATF participation
at reception facilities. The change in academic enrollment, if significant at all, is likely due to
some statewide institutional change (such as a reduction in education funding) since it is relatively
54
4000
6000
8000
1000
012
000
Total
Leve
l 1 P
opula
tion
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13Date
Prisons with Rec. Center Prisons w/out Rec. Center
Showing the impact of the reception adjustmentTime Trend of Level 1 Populations
Figure 10: This figure shows the impact of the reception adjustment on the level 1 subpopulation. Thetime trends are for sum of security level 1 populations parsed by whether the prison has a reception centerfacility or not. The vertical line denotes the last observation prior to implementation of AB 109.
010
000
2000
030
000
Total
Leve
l 2 P
opula
tion
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13Date
Prisons with Rec. Center Prisons w/out Rec. Center
Showing the impact of the reception adjustmentTime Trend of Level 2 Populations
Figure 11: This figure shows the impact of the reception adjustment on the level 2 subpopulation. Thetime trends are for sum of security level 2 populations parsed by whether the prison has a reception centerfacility or not. Note that the effect of RA is conflated with that of AB 109 for this subpopulation. Thevertical line denotes the last observation prior to implementation of AB 109.
55
5000
1000
015
000
2000
0To
tal Le
vel 4
Pop
ulatio
n
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13Date
Prisons with Rec. Center Prisons w/out Rec. Center
Showing the impact of the reception adjustmentTime Trend of Level 4 Populations
Figure 12: This figure shows the impact of the reception adjustment on the level 4 subpopulation. Thetime trends are for sum of security level 4 populations parsed by whether the prison has a reception centerfacility or not. The vertical line denotes the last observation prior to implementation of AB 109.
5000
1000
015
000
2000
025
000
Total
Spe
cial N
eeds
Pop
ulatio
n
Sep10 Jan11 May11 Sep11 Jan12 May12 Sep12 Jan13Date
Prisons with Rec. Center Prisons w/out Rec. Center
Showing the impact of the reception adjustmentTime Trend of Special Needs Populations
Figure 13: This figure shows the impact of the reception adjustment on the special needs subpopulation.The time trends are for sum of special needs populations parsed by whether the prison has a reception centerfacility or not. The vertical line denotes the last observation prior to implementation of AB 109.
56
-.50
.51
Perce
nt Ch
ange
in D
iscipl
inarie
s
Assaults Murder Conduct Drugs Cellphone Mental Total
Changes are averaged for the six month period Jan12 - Jun12 relative to Apr11 - Sep11.
using 6 month averages for main DD treatments9 Month Change in Disciplinary Reports Post AB 109
Base Level IIReception
Figure 14: This figure shows changes in the rate (per 100 inmates) of several types of disciplinaries. Thechanges are in the six month average measured from January 2012 through June 2012, relative to the averagejust prior to implementation of AB 109, April 2011 through September 2011. Source: Generated from CDCRCompStat reporting data.
uniform across the three types of prisons. The increase in SATF beds can be partially explained
by the population decrease, but could imply increased demand for substance abuse treatment since
the waitlist for the program increases quite a bit as well33.
Return to Section 6
33The curious null effect for both SATF categories in the ‘Others’ group of prisons is due to the fact that allSubstance Abuse Treatment facilities are in prisons with a reception center or high proportion of security level IIinmates.
57
0.5
1Pe
rcent
Chan
ge in
Rate
of E
nroll
ment
Academic Work PIA SATF Beds SATF Wait
Enrollment changes are for the six month period Jan12 - Jun12 relative to Apr11 - Sep11.
by main DD treatment groupsPost AB 109 Changes in Average Program Enrollment
Base Level IIReception
Figure 15: This figure shows changes in the rate (per 100 inmates) of several types of program enrollment.The changes are in the six month average measured from January 2012 through June 2012, relative toaverage just prior to implementation of AB 109, April 2011 through September 2011. Source: Generatedfrom CDCR CompStat reporting data.
58